Reported by @oak coyote
Creating a complicated scene with Dall-E needs a way to specify accurately the origin of the ligh, the orientation of the object and the position of the camera. Ideally an engineer plan with x y z coordinate system; instead of vague text indications like "more to the left". Otherwise we enter in an endless cycle of corrections where hallucinations bring us further and further away from the desired result.
a) In this creation, I wanted six different pyramids aligned in a row. Which implies that they each have the same perspective, same angle of view.
b ) In a different example, I had an already constructed landscape with well positionned elements.
a) Each pyramid had a different perspective, making manual correction impossible.
(I have images, but how to add them here?)
b) instead of rotating the whole scene as instructed, Dall-E mixes the elements and place them at random, making a different landscape.
Dall-E via the most recent GPT, windows 11, Brave web browser